Artificial Personal Inference — An API of Oneself Embedded into AI

Original Post on Medium: Artificial Personal Inference — An API of Oneself Embedded into AI | by Sam Bobo | Feb, 2024 | Medium

Intricately detailed, word by word, sentences are crafted on a page. Sentences quickly become paragraphs, and paragraphs become messages. Messages can take the form of emails, books, song lyrics, essays, and like. For the author, each word is carefully chosen, imbued with tone, emotion, and meaning, conveying a level of sophistication and understanding. Diction is unique to the individual, situation, and purpose, ever-evolving with one’s vocabulary and shaped by ones own experience.

Users are often prompted with a series of categories, ranging from tone, format, and length. Inclusive within these categories are subcategories that try to encapsulate a particular set of words or constructs that form the building blocks of a message. Unfortunately, over-generalization removes that unique diction that creates a one-of-a-kind voice that is genuine and yours.

Take, for example, crafting a prompt for a Generative AI system of choice to draft a blog post speaking to the use cases of Artificial Intelligence within Education. The system will generate a summary, as requested, and even incorporate a predefined tonal category, yet, is that diction truly unique to you?

As a nascent blogger of nearly 1 year, I’ve started to tune my attention to word choice and the underlying tone of my blog posts. I read analysts such as Ben Thompson and grasp the types of words he pens on paper, even starting to incorporate into my stylistic choices writing analytical blogs (as an example, many of my titles resemble his style). Yes, one could prompt a generative system to write in the style of Ben, or to create a “Taylor Swift” like song, but those are quickly coming under scrutiny by copyright experts and rightfully so. The ability to generate that type of diction, however, derives from the vast training data ingested within the LLM model in order to probabilistically generate language based on their style.

Shifting my focus back to myself, more broadly, an individual person, the ability to replicate such a voice seizes to exist within AI systems at large. I’ve long claimed that Generative AI solves the Blank Canvas problem, which, in part, it does, however, fails to augment initial drafts with the character that is yourself and exemplify your writing style as to lower the iterations required to finalize a message. For organizations aiming to infuse productivity with AI, such as Microsoft who is poised to win a major segment of the Generative AI sector, a personal voice is something that is missing from Generative AI. This, furthermore, can extend to social media and other creative outlets where brief content (sans entire books where character arcs and fictitious worlds are strategically planned and likely would employ AI systems in an iterative approach). What I am proposing here is a guide for future products, not criticizing the lack thereof, as I understand there are large hurdles to overcome.

In order to build an API for oneself, an Artificial Personal Inference system, one would need to follow the 3 Layers of Artificial Intelligence Ground Truth:

1. Foundation Language Layer — employing the use of Large Language models for the breadth of vocabulary, writing styles, terms of art, patterns used within messages. This can be accomplished today with any LLM, open source or proprietary.
2. Industry Domain Models — instead of an industry, the domain would be yourself. Take any large organization where you are a knowledge worker. Ponder how many emails, word documents, meeting transcriptions, and other various forms of messaging are crafted and disseminated by you on a daily basis. Many large productivity platforms typically include a graphical layer behind the scenes that helps with content discovery and recommendations. That, in theory, could be mined to build an domain model of you.
3. Use Case Layer — This is typically where proprietary data comes into play, specifically, the unique data and expertise of the organization building and deploying the AI system. In the case of an individual, that would be their unique skillset in a particular craft, nuanced methodology (or best practice) of employing that craft and other differentiating features. The approach here would be to utilize few-shot learning to provide the system with prior examples to then produce new results based on the prompt.

An Artificial Personal Inference sounds extremely promising, but why has this not been implemented?

• Tenancy and compute — traditional recommendation systems today rely simply on feature engineering and a scoring algorithm, where parameters can be attuned to a particular individual and stored in a database. Presumably, the same could be done with an LLM but via a vector database, however, reshaping an embedding each time a message is crafted is computationally expensive and more expensive to host. Furthermore, retrieval of said embedding in a multi-tenant environment could prove challenging from a latency, and thus, a use experience perspective.
• Lack of Regulation — with deep fakes, copyright, and fear of losing a job circulate the social sphere around Artificial Intelligence, there is yet a protective mechanism to ensure that an Artificial Personal Inference model would not be used for malintent; couple that with zero-shot text-to-speech synthesis models to completely recreate a person in Artificial Intelligence, both extremely powerful and with the potential for harm.
• Simplistic Interface — many of today’s Generative AI systems incorporated into productivity tools adhere to the cumbersome and overly in debt (technical debt) user interfaces of the past. New modalities of working need to be designed to encourage use and break through in new innovative ways. I outline that in the “Disruption of Design.”
• Context Windows — While the recent announcement of Google’s Gemini 1.5 of a 1M context window is quite astounding and may negate this point in the long run, current models have limited context windows that do not allow for large-scale n-shot prompting which could expedite the creation of said models.

The future ahead for Artificial Intelligence is extremely bright. I recently read an article by Nick Potkalitsky, PhD in Educating AI about his implementation of Artificial Intelligence within the classroom for essay writing with his students. The blog post focused on using Generative AI to break down thesis prompts into sub-prompts (not in the GenAI sense) for students to answer and help form a unique thesis to iterate on. Thereafter, the students continue reading, post-hypothesis thesis, and refine their thesis for an essay over time with the help of GenAI. This type of iterative approach could be used with the Artificial Personal Inference type of model for knowledge workers, creators, and the like to create unique original ideas in partnership with machines.

The use of Artificial Intelligence has immense potential, specifically when focusing on the aspect of scaling human expertise. The moonshot of AI has always been the ability for machines to detect patterns at scale that humans simply could not based on the ability to process sheer amounts of data, say, to solve cancer and the like. While I personally think people are distracted by this goal for other tangential goals like Artificial General Intelligence (AGI), maybe a step in the correct direction will be to augment oneself and then think about the collective whole, or maybe employ the leader-agent model to solve these complex issues.

I still remain a firm believer and dreamer of AI and hope others join me. Lets see what the future holds together!